Optimising TCP over cognitive radio networks for
trains
Kamal Deep Singh, Jean-Marie Bonnin and David Ros
TELECOM Bretagne, Rue de la Châtaigneraie, 35576 Cesson Sévigné cedex, France
Email: {kamal.singh, jmbonnin, david.ros}@telecom-bretagne.eu
Abstract—The Transmission Control Protocol (TCP) is a very
popular transport protocol used by the majority of applications.
However TCP is inefficient when used over mobile networks.
Moreover mobile networks technology is moving towards a
new emerging paradigm called Cognitive Radio (CR). Cognitive
Radio brings its own share of problems for TCP however more
importantly it brings the capability to gather more knowledge
about its radio environment. TCP will face same problems, if
not more, when used over mobile networks enabled with CR
technology. However, these new CR capabilities open the door
for better cross layer TCP optimisation techniques. This paper
proposes a cross layer distributed mechanism using CR triggers
and freeze TCP to optimise TCP over cognitive radio networks
for trains. The results show that our mechanism can improve the
link utilisation efficiency as compared to standard TCP.
Index Terms—TCP, Congestion control, Cognitive radio
I. I NTRODUCTION
Today, the Transmission Control Protocol (TCP) carries the
vast majority of users’ traffic, both in wired and wireless
networks. TCP transports more than 90% of the data over
the Internet, and it is used by applications as diverse as the
web, electronic mail, audio and video streaming (e.g., Spotify,
YouTube), instant messaging, peer-to-peer file sharing, and
even audio- and video-conferencing (e.g., Skype). Moreover,
TCP’s share of Internet traffic has been remarkably stable over
the last decade, despite changes in applications’ popularity.
TCP has proved to be very robust in the face of changes
in networking technology; indeed, TCP was designed in the
1980’s, yet several of its most fundamental features have
changed little since. However, it is well known that TCP
performance - and, thus, the performance of applications
running over TCP - can be very sensitive to the characteristics
of the underlying communication links.
Many proposals have been made to improve TCP in general,
and to better adapt it to some specific link technologies. In
particular, there have been numerous studies of the behavior
of TCP over wireless networks. However, cognitive-radio (CR)
networks and dynamic spectrum allocation techniques present
a new set of challenges to TCP, and in general to layer 4
protocols. An open question is whether existing TCP versions
and enhancements may work well in a CR setting.
In this paper, a cross layer distributed mechanism using CR
triggers and Freeze TCP [5] to optimise TCP over cognitive
This work was supported by the CORRIDOR project (ANR-11-VPTT-004)
of the French national research organisation (ANR).
radio networks for trains is proposed. Freeze TCP freezes the
TCP sender and its timers by advertising a window size of zero
to avoid TCP retransmission timer timeout during temporary
disconnections or network problems. This enables Freeze TCP
to avoid performance degradation due to problems caused
by mobility. This paper is organised as follows. Section II
discusses background and related work. Section III proposes
the mechanism to optimise TCP over cognitive radio networks
for trains. Performance evaluation of our proposal is done in
Section IV and Section V concludes this paper.
II. R ELATED W ORK AND BACKGROUND
A. Cognitive radio
Cognitive radio (CR) [7] is an emerging paradigm for
wireless communications and wireless networking, which is
being touted as the solution to the above problem of spectrum
scarcity. The basic idea of CR-based communications is to
use the radio spectrum in an opportunistic manner, while
respecting some “etiquette” rules to share the spectrum. CRenabled devices constantly monitor their radio environment in
order to find so-called spectrum holes, that is, frequency bands
which are unused at a particular time and location. Such bands
are then employed for communication between CR devices.
Since CR users are likely non-licensed users, they have also
to constantly monitor the used band to detect any activity by
primary users (i.e., licensed users); as soon as primary users
“appear”, CR users have to take measures to avoid interfering
with them, like e.g. moving to a different, free frequency band
or reducing their transmission power.
Cognitive radio is starting to leave the realms of both policy
discussions (e.g., the debate on “TV white spaces”) and academic research, to go into actual applications. This is attested
by related activities in standardization forums like IEEE and
ETSI; both recently-completed and draft industry standards,
namely: IEEE 802.16h [10] and IEEE 802.11af [9], aim at introducing CR features into wireless regional, metropolitan and
local-area networks, respectively. It is therefore of the utmost
importance to understand the performance of CR systems, and
especially the impact they may have on the performance of
users’ applications and users’ Quality of Experience (QoE).
This is especially so since the bulk of the research done in
this space has been devoted to the workings of CR systems
themselves, with little regard for the users’ applications and
protocols running on top of them.
Cognitive radio can be regarded as a natural extension of
software-defined radio (SDR), where a so-called cognition
cycle allows for wireless terminals to be “aware” of their radio
environment and to adapt to it accordingly [11]. Besides some
issues shared by all wireless technologies (like e.g. frequent
packet losses not due to congestion), new features proper to
CR networks like spectrum sensing and spectrum mobility
(i.e., switching to a different spectrum band) may have a strong
impact on the upper layers of the protocol stack.
B. TCP over Cognitive Radio networks
TCP is inefficient when used in mobile networks [1] and
more so in CR networks for the following reasons [3]:
•
•
•
Channels with variable characteristics: CR-enabled devices constantly monitor their radio environment to find
spectrum holes that can be used. However each channel
can have different characteristics in terms of bandwidth,
loss and delay. TCP can be slow to adapt to frequent
changes in the bandwidth and is sensitive to loss and
delay leading to efficiency deterioration.
Temporary disconnections due to spectrum sensing: Devices with CR capabilities alternate between sensing
mode and transmission mode. During sensing mode,
the devices do not transmit the data and this can lead
to sudden increase in round trip time (rtt). If TCP’s
RTO timer TRT O < rtt + To then timeouts will occur
during spectrum sensing period and TCP will set its
congestion window (cwnd) to 1 and slow start threshold
(ssthresh) is reduced to half of the previous value and
TCP enters into the slow-start phase. This leads to the
under-utilisation of the available bandwidth as shown in
[4].
Spectrum handover when primary users appear: Whenever a primary user appears in a channel then secondary
users have to vacate the channel and have to find another
channel to use. In order to transit from one channel to
another, the delay and the disconnection period involved
can degrade TCP performance and also cause TCP round
trip timer timer to expire reducing its efficiency.
C. Related Work
Since its introduction by Mitola [11], the concept of cognitive radio has been the subject of much research effort. The
vast majority of the published studies and proposals are related
to the physical layer (with topics like primary-user detection
being the focus of a good deal of papers), as well as to
the link layer. Moreover, some work has also been devoted
to network-layer. However, to the best of our knowledge,
the relation and interactions between cognitive radio and the
upper protocol layers have not received much attention by
the research community. In particular, there is little published
work on transport-layer issues in a cognitive radio setting, with
only a couple of papers [12], [13] considering the multihop
case at all, and the other studies focusing on simplistic singlehop scenarios [3], [14]. Besides, there exist just a handful
of proposals dealing with improving transport protocol performance, either by means of cross-layer optimization [15],
by optimizing link layer parameters such as frame size for
cognitive radio networks [16], through new protocol proposals
[2] or by integrating an existing protocol into a cognitive
framework [17]. It is however crucial to understand how such
end-to-end protocols and mechanisms would work over CR
systems.
TCP problems in CR networks are also discussed in [6].
Some receiver based TCP improvement solutions to tackle
problems due to wireless conditions and mobility are discussed
in [8] such as:
•
•
DelAck: An ACK for every d packet or send an ACK
if a segment has been unacknowledged for more than T
seconds (eg. 0.1s)
TCP-ADA (TCP adaptive delay acknowledgment) postpones ACKs for a time period in proportion to exponential moving average of inter arrival time between
successive segments. Deferment timer is restarted every
time a new segment arrives before timeout. But can lead
to loss of ACK clocking.
Besides, some proposals for coping with transport-layer
issues in wireless networks could perhaps be useful in CR
networks - such as the work in [5] that proposes an end-toend way to freeze TCP source to tackle disconnections due
to mobility and handover. It is however crucial to understand
how such end-to-end protocols and mechanisms would work
over CR systems.
III. O PTIMISING TCP OVER C OGNITIVE RADIO
NETWORKS FOR TRAINS
In order to improve TCP over Cognitive radio network for
trains, this paper proposes a generic and practical solution
using cross layer techniques. The goal is to improve TCP’s
performance when faced with mobility and wireless conditions. Moreover this optimisation should ensure that there is
no need to modify all the routers and hosts in the Internet.
For this the new solution should be perfectly compatible with
standard TCP.
This paper proposes to implement CR capabilities in the mobile router that will provide connectivity to the clients present
in the train. These CR capabilities will allow the Mobile router
to predict a disconnection and the link layer will send triggers
warning the clients about a pending disconnection as shown
in Figure 1.
A small connection manager application will be installed
in the clients that will reside in the application layer. This
application will also manage TCP connections and our solution
will not modify anything in the TCP stack. It is already a
common practice to install a small software in the clients
before Internet connectivity is provided to them, for example,
when using 3G USB keys. Thus, installing this application
will not degrade the user experience.
TCP
Sender
TCP
Receiver
Mobile
Router
Data packets
Link going
down CR
trigger
RTTlocal
ZWA
TU
advance
Link going
up CR trigger
RTT
sender
TR-ACKS
TD
Disconnection
advance
Fig. 1.
CR triggers sent by Mobile router.
Fig. 2.
Delay diagram for CR triggers.
A. CR triggers
There are two types of CR triggers that are needed from the
mobile router. The first one warns the clients connected to the
mobile router about a pending disconnection and the second
one informs the clients that the link will start working.
•
•
Link going down
Link going up
For the downlink traffic, on the reception of “Link going
down” trigger, the connection manager sends the freeze signal
that sets the receiver window to be zero and thus freezes the
TCP source. On receiving “Link going up” it sends triplicate
ACKs that unfreeze the TCP source. Freeze TCP uses triplicate
ACKs to avoid sender being idle even after the link is up.
For more details please see [5]. For the uplink traffic, the
connection manager directly freezes or unfreezes the TCP
source on the reception of the above respective triggers.
Detection of disconnection start and duration is needed
in advance such that CR triggers can arrive in advance. If
the first trigger is late then the link will go down before the
connection manager can react and if the trigger is early then
the TCP source will freeze and stop sending data even when
the link is available. Both the cases will not be optimal for
TCP performance. As shown in Figure 2 the triggers have
to arrive at least T Uadvance time before a link goes down.
From Figure 2, and as shown later in section IV, it can be
seen that optimal value of T Uadvance is given as follows:
T Uadvance = RT Tlocal + RT Tsender
= RT TT CP .
The trigger “Link going up” has to be sent in advance
too such that when the link is up the triplicate ACKs should
arrive to the TCP source as soon as possible. From Figure 2
it can be seen any value of T Dadvance such that:
RT Tlocal < T Dadvance < RT Tlocal + D
with D as the disconnection duration, will be optimal if the
mobile router can store the arriving triplicate ACKs during
disconnection and later send them as soon as the link is up.
Thus, Mobile router needs to know the values of RT TT CP
and RT Tlocal . Moreover, these values will be different for
different clients connected to it. It will obtain the values
of RT TT CP values for all the clients from the connection
manager installed in those clients as each connection manager
installed on the clients monitors the TCP connections of that
client. The connection manager will monitor the value of
RT TT CP and will send it in response as Mobile router will
periodically request for this value. The RT Tlocal values are
estimated by Mobile router by noting the time elapsed between
sending a request to the connection manager and the arrival
of its response. It should be noted that frequency of such
requests should be low to avoid significant overhead. After
that Mobile router will use the maximum values of RT TT CP
and RT Tlocal among different clients to calculate the values
of T Uadvance and T Dadvance respectively.
Detection of disconnections and re-connections is done
in two ways. Train itineraries are fixed and disconnections
are predicted based on geolocalisation and recorded statistics
during the previous journeys. In addition CR capabilities based
on tracking the signal strength, such as in [19], are used to
predict the pending disconnection. For that this paper proposes
to implement these above two functions in the mobile router.
Freeze TCP
TCP
0
Fig. 3.
TCP Congestion Window
TCP Congestion Window
180
160
140
120
100
80
60
40
20
0
20
40
60
80 100 120 140 160 180 200
Time (s)
Congestion window evolution for Freeze TCP and TCP New Reno.
B. Connection manager
Connection manager software, similar to [18], is installed
on a client when it connects to Mobile router for the first
time. It does not modify the TCP stack, but it keeps a track of
outgoing and incoming TCP packets. When it receives the
“Link going down” trigger from Mobile router it sets the
receiver window in outgoing ACKs as zero, or zero window
advertisement (ZWA), to send the freeze signal to the TCP
source. Similarly when it receives the “Link going up” trigger
it generates triplicate ACKs (TR-ACKs) to unfreeze the TCP
source. The idea is to use advance and precise information on
link going up and down using cognitive radio capabilities.
This approach of tracking TCP flows at the client instead
of tracking them at Mobile router has the advantage that the
approach becomes distributed. Thus, Mobile router doesn’t
have to track many simultaneous connections passing through
it which would have increased the computational load on it.
Rather it just keeps the maximum values of RT TT CP and
RT Tlocal . As discussed in section III-A, connection manager
also monitors the value of RT TT CP for each connection and
sends their values when mobile router requests for them.
IV. R ESULTS
In this section, we study the performance of our proposed
scheme using CR triggers and Freeze TCP over cognitive
radio networks for trains. The study is done using NS-2. The
topology is as shown in Figure 1, the delay between TCP
sources and Node B is 50ms, the delay between NodeB and
Mobile router is 30ms and the delay between Mobile router
and mobile clients present in the train is 10ms. The link
between NodeB and Mobile router has 2Mbps of bandwidth.
The default number of mobile clients is 10 unless specified.
Figure 5 shows that the efficiency of link utilisation is best
when T Uadvance = RT TT CP . This was also discussed before
in section III. Hence from now onwards let us fix the value
of T Uadvance to RT TT CP for next simulations.
Disconnections due to handovers and due to train passing
through tunnels are simulated using an exponential distribution
with average disconnection time of 5 seconds and average
up time of 20 seconds. Figure 3 shows the evolution of
congestion window of Freeze TCP as compared to normal
TCP (New Reno) and for this simulation only 1 mobile
90
80
70
60
50
40
30
20
10
0
Freeze TCP
TCP
0
20
40
60
80 100 120 140 160 180 200
Time (s)
Fig. 4. TCP congestion window evolution when sensing duration To is 0.3
seconds.
client is connected to the mobile router. It can be seen
that congestion window of normal TCP falls every time
there is a disconnection and TCP enters into slow start that
in turn reduces the link utilisation efficiency that is defined as:
ef f iciency =
T otalBitsT ransmitted
Linkuptime∗LinkCapacity
However when Mobile router with CR trigger mechanism
is used then the TCP source freezes the congestion window
as soon as it receives ZWA and later resumes using the same
frozen value of congestion window when disconnection gets
over and the efficiency is improved as compared to normal
TCP as shown in Figure 6.
Moreover, as discussed before, devices with CR capabilities alternate between sensing mode and transmission mode.
During sensing mode, the devices do not transmit the data and
this can introduce spikes in round trip times of TCP packets
leading to TCP’s RTO timer timeouts when sensing duration
To is long for example 0.3 seconds. Figure 7 shows the effect
of increasing values of sensing durations. The value of the
duration of transmission mode is 2 seconds. The results for
To of up to 0.5 seconds are shown as To of the order of
0.3 seconds is considered in [2]. It can be seen that total
throughput over a link having the capacity of 2Mbps decreases
with increasing To . Our scheme that uses CR triggers and
freeze TCP mechanism again shows significant improvement
as compared to standard TCP as it freezes TCP source before
sensing starts and avoids RTO timeouts. Figure 4 shows the
evolution of congestion window for To = 0.3s and it should
be noted that for the first 10 seconds the sensing simulation
is not started in order to allow TCP to finish the initial slow
start phase. As discussed before, with normal TCP, the RTO
timer expires when TRT O < rtt + To and this happens
more frequently for higher values of to thus explaining the
degradation of link utilisation efficiency with normal TCP.
Whereas relatively better performance can be observed with
our scheme that uses CR-triggers and freeze TCP.
V. C ONCLUSIONS
In this paper, a mechanism is proposed that uses link
layer triggers from cognitive radio modules and freeze TCP
to optimise the performance of TCP over cognitive radio
Throughput (Mbps)
Freeze TCP
TCP
0.94
Efficiency
0.92
0.9
0.88
0.86
1.7
1.6
1.5
1.4
1.3
1.2
1.1
1
Freeze TCP
TCP
0
0.84
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
To Sensing Duration (s)
0.82
Fig. 7.
0.8
0.4
0.6
0.8
1
1.2 1.4
TUadvance (RTTs)
1.6
1.8
Fig. 5. Efficiency of Link utilisation with respect to different values of
T Uadvance as a multiple of RTTs.
1
Freeze TCP
TCP
0.9
Efficiency
0.8
0.7
0.6
0.5
0.4
0.3
0
Fig. 6.
2
4
6
8 10 12 14
Number of TCP flows
16
18
Throughput with respect to different values of sensing duration To .
2
20
Efficiency with respect to increasing number of TCP flows.
networks for train. The proposed mechanism uses a cross
layer distributed approach in which a connection manager is
installed on clients wanting to use Internet connectivity on
trains. The connection manager manages and optimises the
TCP connection without changing TCP stack and without
breaking compatibility with standard TCP implementations.
The performance of the proposed mechanism was evaluated
and it showed significant improvement in terms of link utilisation efficiency as compared to standard TCP. In future we
would like to do develop our simulator with more details
related to high speed mobility and additional cognitive radio
parameters.
R EFERENCES
[1] Hari Balakrishnan, Venkata N. Padmanabhan, Srinivasan Seshan, Randy
H. Katz: “A comparison of mechanisms for improving TCP performance
over wireless links.” IEEE/ACM Trans. Netw. 5(6): 756-769 (1997).
[2] K. R. Chowdhury, M. D. Felice and I. F. Akyildiz. “TP-CRAHN: A
Transport Protocol for Cognitive Radio Ad-hoc Networks.” Proc. IEEE,
INFOCOM 2009, Rio de Janeiro, Brazil, Apr. 2009.
[3] Y. R. KondaReddy and P. Agrawal. “Effect of Dynamic Spectrum Access
on Transport Control Protocol Performance.” IEEE GLOBECOM 2009,
Nov. 30 - Dec. 4 2009.
[4] Dilip Sarkar and Harendra Narayan. “Transport Layer Protocols for
Cognitive Networks.”, IEEE Conference on Computer Communications
Workshops, INFOCOM Workshops , 2010.
[5] T. Goff, J. Moronski, D. S. Phatak and V. Gupta. “Freeze-TCP: A
true end-to-end TCP enhancement mechanism for mobile environments”
Proc. IEEE, INFOCOM 2000, Tel-Aviv, Israel, Mar. 2000.
[6] Ian F. Akyildiz, Won-Yeol Lee, Mehmet C. Vuran and Shantidev
Mohanty, “NeXt generation/dynamic spectrum access/cognitive Radio
Wireless Networks: A Survey”. COMPUTER NETWORKS JOURNAL
(ELSEVIER), vol. 50, pp. 2127-2159, 2006.
[7] S. Haykin, “Cognitive radio: brain-empowered wireless communications”, Selected Areas in Communications, IEEE Journal on, Vol. 23,
No. 2. (07 February 2005), pp. 201-220.
[8] Ka-Cheong Leung, Victor O. K. Li, “Transmission control protocol
(TCP) in wireless networks: issues, approaches, and challenges”, IEEE
Communications Surveys & Tutorials, Vol. 8, No. 4. (4th Quarter 2006),
pp. 64-79.
[9] “IEEE P802.11 Task Group af - Wireless LAN in the TV White Space.”
[10] “IEEE Std 802.16h-Part 16: Air Interface for BroadbandWireless Access Systems; Amendment 2: Improved Coexistence Mechanisms for
License-Exempt Operation”, July 2010.
[11] J. Mitola III and G.Q. Maguire Jr. “Cognitive radio: making software
radios more personal.” IEEE Personal Communications, 6(4):13-18,
August 1999.
[12] M. Di Felice, K.R. Chowdhury, and L. Bononi. “Modeling and performance evaluation of Transmission Control Protocol over cognitive
radio ad hoc networks.” In Proceedings of ACM MSWiM, pages 4-12,
Tenerife, October 2009.
[13] M. Di Felice, K.R. Chowdhury, W. Kim, and A. Kassler. “End-to-end
protocols for cognitive radio ad hoc networks: An evaluation study.”
Performance Evaluation, in print, 2011.
[14] T. Issariyakul, L.S. Pillutla, and V. Krishnamurthy. “Tuning radio
resource in an overlay cognitive radio network for TCP: Greed isn’t
good. IEEE Communications Magazine,” 47(7):57-63, July 2009.
[15] C. Luo, F.R. Yu, H. Ji, and V.C.M. Leung. “Optimal channel access
for TCP performance improvement in cognitive radio networks: A
cross-layer design approach.” In Proceedings of IEEE GLOBECOM,
Honolulu, December 2009.
[16] C. Luo, F.R. Yu, H. Ji, and V.C.M. Leung. “Cross-layer design for
TCP performance improvement in cognitive radio networks.” IEEE
Transactions on Vehicular Technology, 59(5):2485-2495, June 2010.
[17] B.S. Manoj, R.R. Rao, and M. Zorzi. “CogNet: a cognitive complete
knowledge network system.” IEEE Wireless Communications, 15(6):8188, December 2008.
[18] Rayene Ben Rayana, J-M Bonnin, “Mobility aware application manager
for mobile networks.” ITST 2008. 8th International Conference on ITS
Telecommunications, Pages 337-342.
[19] Yoo, Sang-Jo, David Cypher and Nada Golmie, “LMS Predictive Link
Triggering for Seamless Handovers in Heterogeneous Wireless Networks.” Military Communications Conference, 2007, IEEE MILCOM
2007, 29-31 Oct.